A novel method for leakage monitoring in Network-Level urban medium- and Low-Pressure natural gas pipelines combining information theory and Light Gradient Boosting

Abstract

Leaks in urban medium- and low-pressure natural gas pipeline networks pose substantial risks and detection difficulties, compromising pipeline network reliability and urban safety. Little research has been conducted on leakage monitoring of network-level urban pipelines. This paper proposes a machine learning framework based on a dataset consisting of 141,236 samples collected over nearly five years. The framework real-time classifies the causes of anomalous signals (with leaks being one of the causes) collected by IoT terminals located near each section of pipelines within a large network at both the individual-sampling-point level and entire-incident level, thus enabling monitoring. Feature extraction is a crucial part of machine learning, relying solely on the sensor data and computational rules defined through diffusion laws and data analysis, requiring no prior information. The original feature set includes 360 features with physical significance. A novel iMICRS reduction method integrating information theory and rough set theory is developed to determine the optimal feature combination. Combining ten-fold cross-validation and Bayesian optimization, LightGBM achieves the highest ROC_AUC of 0.871 on a test set covering 9,885 sampling points. SHAP is used for prediction interpretation. The classification method for the entire incident based on the prediction results achieves a recall rate of 90% in diagnosing leaks (including multiple small leaks). This study provides an effective full-process engineering solution for leakage monitoring in urban medium- and low-pressure pipeline networks, based on actual operational data.